X-STARS

This is the official PyTorch implementation of the "Cross-sensor self-supervised training and alignment for remote sensing" paper, published on JSTARS.

This repository builds upon the original DINO implementation. You can follow that repo to install the required packages.

Pretraining

To run the pretraining from scratch, you can run:

python pretraining.py --arch vit_tiny --data_path data/path/ --output_dir /output/directory --epochs 400 --batch_size_per_gpu 4 --use_msad --msad_embedding_dim 192 --sensors Sentinel Landsat --mean 0.15590523 0.15850738 0.10111853 --std 0.14238988 0.11567883 0.0910672

To run the continual pretraining, you can run:

python continual_pretraining.py --arch vit_tiny --data_path data/path/ --output_dir /output/directory --epochs 400 --batch_size_per_gpu 12 --use_msad --msad_embedding_dim 192 --sensors Sentinel Landsat --mean 0.15590523 0.15850738 0.10111853 --std 0.14238988 0.11567883 0.0910672 --adapt_sensor Landsat --pretrained_weights pretrain/net/weights 

Dataset

The dataset class is shaped on the MSC-France dataset, presented in the already mentioned paper. The name of the images is the same for each sensor. The directories are organized as follows:

MSC-France
├─Sentinel
   ├─Bordeaux
   ├─Grenoble
   ...
   └─Toulouse
├─Landsat
   ├─Bordeaux
   ├─Grenoble
   ...
   └─Toulouse
└─SPOT
   ├─Bordeaux
   ├─Grenoble
   ...
   └─Toulouse

Dataset is downloadable here.

Model weights

The pre-trained models are available at this link.

Citation

@article{marsocci2025cross,
  title={Cross-sensor self-supervised training and alignment for remote sensing},
  author={Marsocci, Valerio and Audebert, Nicolas},
  journal={IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing},
  year={2025},
  publisher={IEEE}
}